Research Of SIFT Stereo Matching Algorithm Based On GPU

Currently, GPU has been widely used in general purpose computing. Due tothe programmability of hardware, the applications of earlier general-purposecomputing are limited, the development is also very difficult. NVIDIA launchedthe CUDA, it is a parallel computing device as a data system of hardware andsoftware, the programming style is simple, it uses a multi-threaded parallelprocessing and makes people take good advantage of GPU.On the other hand, image matching is a key technology of the imageprocessing, it is the basis for the more advanced image processing technology.SIFT stereo matching algorithm is a stable feature-based matching algorithm.SIFT has better resistance for scaling, rotation transformation and illumination.It is a hot research field of image processing.While SIFT stereo matching algorithm has many advantages, it is atime-consuming operation, and since it is a matching algorithm based on feature,the number of matching points is less. The obtained matching points can notsatisfy the demand of generating disparity map. Therefore, the applications arelimited. SIFT stereo matching algorithm exists these two problems. On the onehand, the traditional pre-processing algorithms are discussed, this paper analyzesthe parallelism of the algorithms, and implements gray conversion, Gaussianfiltering, histogram equalization, and Wallis filtering on CUDA. And the paperimprove the original SIFT algorithm, it makes matching points that get from theoriginal algorithm as seed points, and then make regional growth, traverse theentire image, so you can get many matching points which are good for generatingthe disparity map; on the other hand, the paper uses the CPU and GPUheterogeneous platforms and analyses the CUDA programming model andmemory model. We divise tasks of SIFT stereo matching algorithm and analysesthe algorithm in detail, so the algorithm can be carried out on CUDA. GPUaccelerates the speed of the algorithm.The experiment of the regional growth algorithm based on SIFT was carriedout on the platform of Binocular vision, it verified the effectiveness of theimproved algorithm. In order to evaluate the acceleration of GPU more directly,the experiment of the SIFT stereo matching algorithm based on GPU was carriedout in a variety of different scenarios, the results show that the algorithm notonly improves the speed of the algorithm on CUDA, but also to ensure thestability of the original algorithm. The efficiency of the algorithm has been greatly improved.